Carlos Gonzalez-Perez

University of Oxford · Economics · 2026–27 job market

Fields

Econometrics, Theory

Job market paper

Adaptive Design of Clustered Surveys and Experiments

When designing clustered surveys or experiments on a budget, researchers must decide between sampling more clusters or more units units per cluster. Unfortunately, the variance-minimizing design depends on the intra-cluster correlation (ICC), a property of the data which has not yet been collected. Therefore, practitioners are generally forced to guess this parameter. I show that guessing is first-order expensive: a design built on a misspecified ICC forfeits precision of the same order as the target variance itself. Instead, I propose two simple algorithms that learn the ICC while the experiment runs. These algorithms attain negligible excess variance compared to a designer who knows the true ICC, improve an order of magnitude upon any misspecified static design, and are rate optimal. Moreover, adaptivity need not compromise inference. Under a symmetry assumption on the errors, mean estimators remain asymptotically normal conditional on the realized design, with no data discarded and no correction. This inference invariance principle extends beyond the cluster setting and validates standard inference for popular sampling procedures such as Neyman allocations with variances computed from pilot data. Revisiting Fink et al. (2020b), I find that my methods would have lowered the variance of their estimators by up to 41% at equal cost or reduced the cost by up to 42% at equal precision. I further demonstrate the feasibility of the algorithms by using them to field an online survey in real time. The results show that the ICC can be learned at only second-order cost and without compromising inference.